Adaptation Of A Multi-Site Network To A New Clinical Site Via Batch-Normalization Similarity

This paper tackles the challenging problem of medical site adaptation; i.e., learning a model from multi-site source data such that it can be modified and adapted to a new site using only unlabeled data from the new site. The method is based on Domain Specific Batch Normalization architecture and uses the Batch Normalization statistics of the new site to find the most similar internal site. The similarity measure is computed in an embedded space of the BN parameters. We evaluated our method on the task of segmentation of prostate MRI data from six different institutions with distribution shifts acquired from public datasets. The experimental results show that this approach outperforms other generalization and adaptation methods across almost all settings of unseen sites.